Predictive maintenance convolutional neural networks

Abstract

Systems and methods provide techniques for performing predictive maintenance data analysis based on telemetry data. In one embodiments, a method includes at least operations configured to obtain a training telemetry data object, determine a training input data object based on the training telemetry data object, obtain a maintenance data object, and generate a trained predictive maintenance convolutional neural network based on the training input data object and the maintenance data object. The trained predictive maintenance convolutional neural network can be utilized to generate maintenance predictions for a monitored system. The maintenance prediction can include data objects with visual explanatory capabilities, such as data objects that describe heatmaps over input telemetry data.

The invention claimed is: 
     
       1. A computer-implemented method for performing predictive maintenance of a monitored system, the computer-implemented method comprising:
 obtaining, a training telemetry data object corresponding to telemetry data indicative of time-series sensor measurement data collected from a plurality of sensors of the monitored system; 
 determining, based on the training telemetry data object, a training input data object which is in a data format defined by an input structure of a predictive maintenance convolutional neural network, wherein the training input data object comprises one or more training input channels which are based on a type of the training input data object, and wherein each training input channel is associated with a sensor from amongst the plurality of sensors and based on a subset of the telemetry data; 
 obtaining a maintenance data object for the monitored system, wherein the maintenance data object identifies one or more maintenance-critical timestamp of the time-series sensor measurement data; and 
 generating, based on the training input data object and the maintenance data object, a trained predictive maintenance convolutional neural network, wherein the trained predictive maintenance convolutional neural network is configured to process a predictive input data object for the monitored system to generate a maintenance prediction for the monitored system. 
 
     
     
       2. The computer-implemented method of  claim 1 , further comprising:
 obtaining a predictive telemetry data object corresponding to predictive telemetry data indicative of predictive time-series sensor measurement data for the monitored system; 
 determining, based on the predictive telemetry data object, a predictive input data object, wherein the predictive input data object comprises one or more predictive input channels, and wherein each predictive input channel of the one or more predictive input channels is associated with the sensor of the plurality of sensors, and further wherein each predictive input channel of the one or more predictive input channels is determined based on a first-subset of the predictive time-series sensor measurement data that is associated with the sensor for the predictive input channel; and 
 processing the predictive input data object in accordance with the trained predictive maintenance convolutional neural network to generate the maintenance prediction for the monitored system, and wherein the monitored system is a heating, ventilation, and air-conditioning system, and wherein the trained predictive maintenance convolutional neural network is generated using a gradient descent training routine. 
 
     
     
       3. The computer-implemented method of  claim 2 , further comprising:
 for each predictive input channel of the one or more predictive input channels,
 determining a per-input-channel weight value for the predictive input channel with respect to a target classification category associated with the maintenance prediction; and 
 generating a per-input-channel heatmap for the predictive input channel based on the per-input-channel weight value for the predictive input channel and a feature map for the predictive input channel, wherein the per-input-channel heatmap for the predictive input channel indicates a per-component predictive significance value for each component of a plurality of components of the predictive input channel. 
 
 
     
     
       4. The computer-implemented method of  claim 3 , wherein generating the per-input-channel heatmap for the predictive input channel comprises:
 determining a per-input-channel activation output based the per-input-channel weight value for the predictive input channel and the feature map the predictive input channel; and 
 applying an activation function to the per-input-channel activation output to generate the per-input-channel heatmap for the predictive input channel, and wherein the activation function is a rectified linear unit activation function. 
 
     
     
       5. The computer-implemented method of  claim 3 , wherein generating the per-input-channel heatmap for the predictive input channel comprises performing guided backpropagation. 
     
     
       6. An apparatus for performing predictive maintenance of a monitored system, the apparatus comprising at least one processor and at least one memory including program code, the at least one memory and the program code configured to, with the processor, cause the apparatus to at least:
 obtain a training telemetry data object corresponding to telemetry data indicative of time-series sensor measurement data collected from a plurality of sensors of the monitored system; 
 determine, based on the training telemetry data object, a training input data object which is in a data format defined by an input structure of a predictive maintenance convolutional neural network, wherein the training input data object comprises one or more training input channels which are based on a type of the training input data object, and wherein each training input channel is associated with a sensor from amongst the plurality of sensors and based on a subset of the telemetry data; 
 obtain a maintenance data object for the monitored system, wherein the maintenance data object identifies one or more maintenance-critical timestamp of the time-series sensor measurement data; and 
 generate, based on the training input data object and the maintenance data object, a trained predictive maintenance convolutional neural network, wherein the trained predictive maintenance convolutional neural network is configured to process a predictive input data object for the monitored system to generate a maintenance prediction for the monitored system. 
 
     
     
       7. The apparatus of  claim 6 , wherein the program code is further configured to, with the processor, cause the apparatus to at least:
 obtain a predictive telemetry data object corresponding to predictive telemetry data indicative of predictive time-series sensor measurement data for the monitored system; 
 determine, based on the predictive telemetry data object, a predictive input data object, wherein the predictive input data object comprises one or more predictive input channels, and wherein each predictive input channel of the one or more predictive input channels is associated with the sensor of the plurality of sensors, and further wherein each predictive input channel of the one or more predictive input channels is determined based on a first subset of the predictive time-series sensor measurement data that is associated with the sensor for the predictive input channel; and 
 process the predictive input data object in accordance with the trained predictive maintenance convolutional neural network to generate the maintenance prediction for the monitored system. 
 
     
     
       8. The apparatus of  claim 7 , wherein the program code is further configured to, with the processor, cause the apparatus to at least:
 for each predictive input channel of the one or more predictive input channels,
 determine a per-input-channel weight value for the predictive input channel with respect to a target classification category associated with the maintenance prediction; and 
 generate a per-input-channel heatmap for the predictive input channel based on the per-input-channel weight value for the predictive input channel and a feature map for the predictive input channel, wherein the per-input-channel heatmap for the predictive input channel indicates a per-component predictive significance value for each component of a plurality of components of the predictive input channel. 
 
 
     
     
       9. The apparatus of  claim 8 , wherein generating the per-input-channel heatmap for the predictive input channel comprises:
 determining a per-input-channel activation output based the per-input-channel weight value for the predictive input channel and the feature map the predictive input channel; and 
 applying an activation function to the per-input-channel activation output to generate the per-input-channel heatmap for the predictive input channel. 
 
     
     
       10. The apparatus of  claim 8 , wherein generating the per-input-channel heatmap for the predictive input channel comprises performing guided backpropagation. 
     
     
       11. The apparatus of  claim 6 , wherein the monitored system is a heating, ventilation, and air-conditioning system. 
     
     
       12. A computer program product for performing predictive maintenance of a monitored system, the computer program product comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions configured to:
 obtain a training telemetry data object corresponding to telemetry data indicative of time-series sensor measurement data collected from a plurality of sensors of the monitored system; 
 determine, based on the training telemetry data object, a training input data object which is in a data format defined by an input structure of a predictive maintenance convolutional neural network, wherein the training input data object comprises one or more training input channels which are based on a type of the training input data object, and wherein each training input channel is associated with a sensor from amongst the plurality of sensors and based on a subset of the telemetry data; 
 obtain a maintenance data object for the monitored system, wherein the maintenance data object identifies one or more maintenance-critical timestamp of the time-series sensor measurement data; and 
 generate, based on the training input data object and the maintenance data object, a trained predictive maintenance convolutional neural network, wherein the trained predictive maintenance convolutional neural network is configured to process a predictive input data object for the monitored system to generate a maintenance prediction for the monitored system. 
 
     
     
       13. The computer program product of  claim 12 , wherein the computer-readable program code portions are further configured to:
 obtain a predictive telemetry data object corresponding to predictive telemetry data indicative of predictive time-series sensor measurement data associated with the plurality of sensors of the monitored system; 
 determine, based on the predictive telemetry data object, a predictive input data object, wherein the predictive input data object comprises one or more predictive input channels, and wherein each predictive input channel of the one or more predictive input channels is associated with the sensor of the plurality of sensors, and further wherein each predictive input channel of the one or more predictive input channels is determined based on a first subset of the predictive time-series sensor measurement data associated with the sensor for the predictive input channel; and 
 process the predictive input data object in accordance with the trained predictive maintenance convolutional neural network to generate the maintenance prediction for the monitored system. 
 
     
     
       14. The computer program product of  claim 13 , wherein the monitored system is a heating, ventilation, and air-conditioning system. 
     
     
       15. The computer program product of  claim 13 , wherein the trained predictive maintenance convolutional neural network is generated using a gradient descent training routine. 
     
     
       16. A computer-implemented method for performing predictive maintenance of a monitored system, the computer-implemented method comprising:
 obtaining a predictive telemetry data object corresponding to telemetry data indicative of predictive time-series sensor measurement data associated with a plurality of sensors of the monitored system; 
 determining, based on the predictive telemetry data object, a predictive input data object which is in a data format defined by an input structure of a predictive maintenance convolutional neural network, wherein the predictive input data object comprises one or more predictive input channels which are based on a type of the predictive input data object, and wherein each predictive input channel is associated with a sensor from amongst the plurality of sensors and based on a subset of the telemetry data; and 
 processing the predictive input data object in accordance with a trained predictive maintenance convolutional neural network to generate a maintenance prediction for the monitored system. 
 
     
     
       17. The computer-implemented method of  claim 16 , further comprising:
 for each predictive input channel of the one or more predictive input channels,
 determining a per-input-channel weight value for the predictive input channel with respect to a target classification category associated with the maintenance prediction, wherein the maintenance prediction describes at least one of: one or more affected hardware devices of a plurality of existing hardware devices associated with the monitored system and one or more predicted maintenance need dates for the monitored system; and 
 generating a per-input-channel heatmap for the predictive input channel based on the per-input-channel weight value for the predictive input channel and a feature map for the predictive input channel, wherein the per-input-channel heatmap for the predictive input channel indicates a per-component predictive significance value for each component of a plurality of components of the predictive input channel. 
 
 
     
     
       18. The computer-implemented method of  claim 17 , wherein generating the per-input-channel heatmap for the predictive input channel comprises:
 determining a per-input-channel activation output based the per-input-channel weight value for the predictive input channel and the feature map the predictive input channel; and 
 applying an activation function to the per-input-channel activation output to generate the per-input-channel heatmap for the predictive input channel, and wherein the activation function is a rectified linear unit activation function. 
 
     
     
       19. The computer-implemented method of  claim 17 , wherein generating the per-input-channel heatmap for the predictive input channel comprises performing guided backpropagation. 
     
     
       20. The computer-implemented method of  claim 16 , wherein generating the trained predictive maintenance convolutional neural network comprises:
 obtaining a training telemetry data object corresponding to telemetry data indicative of time-series sensor measurement data collected from the plurality of sensors of the monitored system; 
 determining, based on the training telemetry data object, a training input data object, wherein the training input data object comprises one or more training input channels, and wherein each training input channel of the one or more training input channels is associated with the sensor of the plurality of sensors, and further wherein each training input channel of the one or more training input channels is determined based on a first subset of time-series sensor measurement data associated with the sensor for the training input channel; 
 obtaining a maintenance data object for the monitored system, wherein the maintenance data object identifies one or more maintenance-critical timestamp of the time-series sensor measurement data; and 
 generating, based on the training input data object and the maintenance data object, a trained predictive maintenance convolutional neural network.